Overview

➢ Objective and Scope

➢ Literature Survey
➢ Problem Formulation
➢ Experimental results
➢ Conclusion
➢ References
Objective
Our objective in this project is to stimulate and experiment the methods proposed by
previous works in VRPTW that is to minimize the number of vehicles and total distance
traveled to service the customers without violating the capacity and time window
constraints and to explore various areas to get better solutions of The Vehicle Routing
Problem with Time Windows (VRPTW) using multi-objective genetic algorithm to develop
more refined approach by experimenting with different type of crossover and mutation
methodology and rate. A set of well-known benchmark data are used to compare the
effectiveness of the proposed method for solving the VRPTW. The scope of project
include getting optimal solution for upto hundred customers using Multi-Objective
Genetic Algorithm( MOGA) with different number of generations.
Literature Survey
In 1959, VRP first appeared in a paper[3] by George Dantzig and John Ramser formulated as a truck
dispatching problem.

Different variations of VRP has been developed over the years.The problem we are working on is a hybrid
of Capacitated Vehicle Routing Problem (CVRP) and Vehicle Routing Problem with Time Windows
(VRPTW).

Researchers have studied VRPTW using exact and approximation techniques.

Kohl's work [4] is his one of the most efficient and accurate methods of VRPTW. He was able to solve 100
different customer sized instances. However, no algorithm has been developed that can optimally solve
all his VRPTWs with more than 100 customers.
Literature Survey(Contd.)
Gehring and Homberger [5] presented a two-stage hybrid search that first uses an evolutionary strategy to
minimize the number of vehicles and then uses a tabu search algorithm to minimize the total distance.

Luca Maria Gambardella et al.[6] studied one type of multi-objective implementation of VRPTW by minimizing a
hierarchical objective function.Here the first objective was to minimize the number of vehicles and the second is
to minimize the total travel time. This was done by adapting the Ant Colony system "ACS". [7] Define two ant
colonies, each dedicated to the optimization of each objective function.

Ombuki et al. (2006) presented a multi-objective genetic algorithm for the time window problem of vehicle routes,
aiming at minimizing the total number of vehicles and the total cost (distance). When resolving, they used Best
Cost Route Crossover (BCRC) as the crossover operator and Constrained Route Reversal Mutation as the
mutation operator. They concluded that their approach was effective because the solutions obtained competed
with the best-known solutions in the literature.the approach was in two stages: GA was used first to establish the
number of vehicles, and in a local tabu search was conducted to minimize the total cost of traveled
distance.Essentially, the multi-target VRPTW problem turned into a single-target optimization.
Problem Formulation
What is VRPTW(Vehicle Routing Problem with Time Windows)?
Delivering goods to customers at different locations with each customer
having a hard time window and multiple vehicles is used with homogenous
capacity.
Scenarios to be tackled
● When capacity of the vehicle exceeds the maximum limit then number of vehicles should be
increased.
● When a vehicle arrive before ready time it needs to wait till ready time i.e. hard time window is
followed
● When vehicle arrive after due time then a new vehicle is needed and hence distance travelled
increases.
Problem Formulation(Contd.)
This problem is a real time adaptation of Travelling Salesman Problem(TSP).Hence It is a NP
Complete problem and thus a heuristic and metaheuristic approach would be suitable to get optimal
solution.
Problem Formulation(Contd.)
We are working on Multi-Objective Genetic Algorithm approach for this problem.

Genetic Algorithm (GA) is a widely used search method for complex problems such as VRP. Selection, Crossover, and
Mutation are the main operators used in GA.
Problem Formulation(Contd.)
Data Structure Used:
Each customer is represented as a gene and an array of customers that will show the routes of the
vehicles is chromosome. A chromosome will represent a solution for the problem. A group of
chromosomes represented as population.

This example
has 6customers
Problem Formulation(Contd.)
Population Generation

Initial population is generated with population size 100 by selecting random permutation of customers and then GA operators are applied to the population
to get further optimized solution with each iteration as shown in diagram of typical Genetic algorithm above.

Crossover

Following crossover has been experimented:

1.PMX (Partially-Mapped Crossover)- Subscripts P1 and P2 (with P1 and P2) are uniformly chosen here at random. Through position-by-position
exchange procedures, the two points (P1 and P2) define a matching section that is used to make a cross.

The two sub-routes are typically switched between the crossing points via PMX. There are multiple batch numbers. The two new solutions then
undergo a different procedure. If a duplicate batch number in solution 1 is discovered, it is exchanged with the batch number of solution 2 in front of
the first duplicate batch location of solution 1 (for example, solution 1's solution 5 is replaced with solution 2's solution 7). For duplicate solution 2
batch numbers, the same process is used. The procedure is repeated until all duplicates are gone.
PMX Crossover
Problem Formulation(Contd.)
2.Cyclic Crossover: In this crossover, the positions of the genes in both parents are matched. Then a cycle is
created by traversing through the genes in one of the parents. The genes in the other parent that are in the same cycle
are then copied to the offspring in the same positions.

Cyclic crossover can be useful in maintaining the order of genes that are important for the solution quality and can
be applied to various optimization problems such as scheduling, routing, and packing.
Cyclic Crossover
Problem Formulation(Contd.)

3.Best Cost Route Crossover, BCRC: It aims at minimizing the number of vehicles
and cost at the same time while considering feasibility constraints.

In BCRC, two parents are selected randomly lets say P1 and P2, from these two parents
some genes are selected from a route and in opposite parent these genes are deleted.
Now one by one genes deleted from parents are inserted at best possible locations to
create two children C1 and C2 from P1 and P2.
Problem Formulation(Contd.)

BCRC Example
Explanation
Problem Formulation(Contd.)
Mutation
Following mutation has been experimented

1.Swap Mutation: It select two positions in the chromosome at random and interchange the values.

2.Cyclic Mutation: It selects a subset of consecutive genes and cyclically shifts their values .

3.Inversion Mutation: It selects a subset of genes and inverts the entire string in the subset.
Problem Formulation(Contd.)
Selection
After crossover and mutation new children are generated and are added to the population.

An elite model is incorporated to ensure that the best talent is passed on to the next generation. The
advantage of the elitist method over traditional stochastic reproduction is that it ensures that the
current best solution from the previous generation is copied unchanged to the next generation. This
means that the best solution produced by the best overall chromosome never degrades from one
generation to the next.

We sort the population on basis of pareto ranking and size of population is restored to initial
population size. The chromosomes with better pareto ranking are retained and chromosomes with not
so good ranking are eliminated.
Problem Formulation(Contd.)
Problem Formulation(Contd.)
Termination
In GA we run our algorithm to a number of generations till a termination condition is met. We
are taking number of generations as a terminating condition in our case. For now we take
number of generations as 450 for termination condition.
Experimental Results
Execution Environment
The code was written in C++ and executed on a computer with 8GB of RAM and an Intel(R)
Core(TM) i5-8265U, 1.60GHz processor.
Experimental Results(Contd.)
Analysis of Dataset
The typical Solomon's VRPTW benchmark [9] problem examples are used in our experimental findings.

Solomon's data is grouped in six classification: R1,R2,C1,C2,RC1,RC2.

R ➔ Uniformly distributed customer locations.

C ➔ Clients are grouped either geographically or according to time intervals.

RC➔ Hybrid problem with mixed features from both R and C.

Experimental Results(Contd.)
Experiments and Obtained Results
We have experimented with different combinations of crossover and mutation that have the
potential to be used for this scenario i.e. VRPTW problem. The crossover and mutation
explained above is experimented and the results obtained are recorded and shown in next slide
for a different number of generations.
Experimental Results(Contd.)
Experimental Results(Contd.)

Thus we can conclude that swap mutation and inversion mutation are performing better than
cyclic mutation and giving some promising results. PMX is performing better than the Cyclic
crossover operator, as the Cyclic crossover is distorting the routes to more extent.

As BCRC [11] has been introduced specifically for this problem, it is outperforming PMX and
Cyclic crossover. When a crossover is performed on two parents using BCRC the children
produced are either better or equal to parents in terms of cost, since we are comparing them with
their parents for each iteration.
Experimental Results(Contd.)

The experimentation on BCRC and Fitness Function are based on the set of GA settings listed
below:

• population size = 100

• generation span = 450

• crossover rate = 0.80

• mutation rate = 0.10

Experimental Results(Contd.)
We are using the fitness function as a parameter to compare the cost of children to their parents in BCRC crossover and we
are using the Pareto ranking method to select the elitist population for the next iteration.

The fitness function is formulated as

Fitness = α*Number_of_Vehicles + β*Distance.

Increasing the value of α prioritizes the number of vehicles whereas increasing the value of β prioritizes the total distance
traveled.

Different combinations of α and β have been used to obtain the results in Table below.

In every iteration of GA, we are performing crossover and mutation and then selecting the elitist population for the next
generation, but we deliberately add some worse population to the next generation which provides randomness and prevents
GA to stuck at local minima. We are adding a worse population in a small proportion (2% to 10%).
Experimental Results(Contd.)

Comparison with best

published result using
Solomon's benchmark
problems
Experimental Results(Contd.)

We have compared the results of different datasets using BCRC and inversion mutation (with 3
elements to be inverted). ↑ indicates our results are better (either in terms of distance or in
terms of the number of vehicles) than the best-known while ≈ indicating that our results are the
same as the best-known results.
Experimental Results(Contd.)

Route
Configuration
of C101.
Experimental Results(Contd.)

Network configuration for 100

spatially grouped customers
across a large time range. Test out
issue C101.
Experimental Results(Contd.)

Route
Configuration of
R104.
Experimental Results(Contd.)

network configuration for

100 evenly-spaced customers
with a limited time window
Experimental Results(Contd.)

Depending on whether the customer wants the greatest solutions for trip costs or the best
number of vehicles, there are two (or more) solutions when utilizing the Pareto ranking. There
may be a single Pareto solution in some experiments that is best known in both the vehicle and
distance dimensions, such as experiment c101. Other approaches, like r103, shorten the distance
but at the expense of introducing more vehicles.
Conclusion

We have experimented with the different parameters used in the genetic algorithm which
focuses on solving the VRPTW with MOGA. This paper constitutes the experimentation
done with different crossover and mutation operators and the difference between their
results. Our results are comparable to other vehicle-biased results in the literature. We
have used a variation of BCRC that is suggested in [11]. The result from data instance
C101 is the same as the best-known result. And some of the data instances like R103 and
RC102 have shown optimized results in terms of one of the variables i.e. either number of
vehicles or distance.
References
1. M. R. Garey, and D. S. Johnson, Computers and Intractability, A Guide to The Theory of NP-Completeness, W. H. Freeman and Company,
1979.
2. J. K. Lenstra and A. H. G. Rinnooy Kan, “Complexity of Vehicle Routing Problem with Time Windows,” Networks, 11:221-227, 1981.
3. http://andresjaquep.files.wordpress.com/2008/10/2627477-clasico-dantzig.pdf
4. Niklas Kohl. “Exact Methods for Time Constrained Routing and Related Scheduling Problems,” PhD Thesis, Department of Mathematical
Modeling,Technical University of Denmark, 1995.
5. H. Gehring and J. Homberger, “Parallelization of a Two-phased Metaheuristic for Routing Problems with Time Windows, ” Asia- Pacific
Journal of Operational Research 18, 35-47, 2001.
6. L. M, Gambardella, E. Taillard and G. Agazzi, “MACS-VRPTW: A Multiple Ant Colony System for Vehicle Routing Problems with Time
Windows,” In David Corne, Marco Dorigo, and Fred Glover, editors, New Ideas in Optimization, pp.63-76, McGraw-Hill, London, 1999.
7. M. Dorigo, L. M. Gambardella, “Ant Colony System: A Cooperative Learning Approach to the Traveling Salesman Problem,” IEEE
Transactions on Evolutionary Computation 1:53-66, 1997.
8. https://github.com/raunaktrip/Btech_project_Moga_vrptw
9. http://w.cba.neu.edu/~msolomon/problems.htm
10. http://w.cba.neu.edu/~msolomon/r1r2solu.htm
11. https://www.researchgate.net/publication/227309327_Multi-
Objective_Genetic_Algorithms_for_Vehicle_Routing_Problem_with_Time_Windows